AI Case Study
Researchers at Florida State University predict one-year mortality rate in ICU patients suffering from a heart disease
Researchers at Florida State University and the University of Florida have assessed machine learning models to be able to predict one-year mortality rate in patients with Acute Myocardial Infarction and Post Myocardial Infarction Syndrome. The team, that used multiple datasets obtained from MIMIC-III, was able to identify the most accurate machine learning algorithms. Those are models that apply Logistic Model Trees (LMT) and Simple Logistic algorithms.
Industry
Healthcare
Healthcare Providers And Services
Project Overview
"Compared with risk assessment guidelines that require manual calculation of scores, machine learning-based prediction for disease outcomes such as mortality can be utilized to save time and improve prediction accuracy.
The current study built and evaluated various machine learning models to predict one-year mortality in patients that were diagnosed with an AMI or post myocardial syndrome (PMS). One-year mortality was selected for this study as a starting point because it would allow for comparison to other studies and consider that there are patients that had multiple AMI admissions within a two-year period. In our study, the results of the best performing machine learning model are compared to a deep feedforward neural network (Deep FNN) with back propagation. There are some overlapping characteristics compared to previous cardiac associated prediction models and some differences. For example, this study does not differentiate between NSTEMI and STEMI as many previous models do. In addition, this study does not use any interpreted data such as echocardiography, location of AMI, or Killip class. The goal is to build prediction models using different machine learning algorithms and datasets to determine which classification model(s) performs the best."
Reported Results
"One interesting finding in this study, was the importance of initial diagnosis, as a “myocardial infarction” or “rule out myocardial infarction”. Of the 5436 admissions, 3375 of these were not tagged as either of these as the primary reason for admission. Of these, the percentage in the positive instances was higher than the percentage in the positive instances (34.3% versus 30%). Chi-square statistics was applied to compare the expected outcome with the observed outcome. The p-value was < .00001 and the result was significant at p < .05. This indicates the importance of recognizing myocardial infarction initially so that appropriate treatment can be used as soon as possible. Another possible explanation for this is when patient was hospitalized for a comorbidity, a significant number of AMI occurred. As there is no way to time stamp when the AMI occurred for patients who had AMI before admission, it is not possible to differentiate between those that occurred prior to admission and after admission. If that is the case, it indicates that an AMI that occurs during an admission may lead to a higher mortality rate."
Technology
"The results of the best performing shallow prediction models were compared to a deep feedforward neural network (Deep FNN) with back propagation. We included a cohort of 5436 admissions. Six datasets were developed and compared. The models applying Logistic Model Trees (LMT) and Simple Logistic algorithms to the combined dataset resulted in the highest prediction accuracy at 85.12% and the highest AUC at .901."
Function
Background
"Heart disease is the leading cause of death in the United States. About half of Americans have at least one of the three most common risk factors for heart disease: high blood pressure, high low-density lipoprotein (LDL) cholesterol, and smoking. Other factors that contribute to cardiovascular health are physical activity, diet, weight, and glucose control. By 2035, it is estimated that almost half of adults in the United States will have some form of heart disease, with the cost exceeding $1.1 trillion."
Benefits
Data
5436 admissions